Internet of Sounds: A Distributed Audio Paradigm
- Internet of Sounds (IoS) is a paradigm that integrates sound sensing, processing, indexing, and actuation across networked devices for diverse audio applications.
- It employs a multi-layered architecture combining physical sensors, core processing methods, and application-specific modules to enable real-time sound analytics.
- IoS leverages distributed computing frameworks, such as cloud-fog-device configurations, to achieve scalable, low-latency, and adaptable sound monitoring and augmentation.
Internet of Sounds (IoS) is an emergent paradigm in which sound sensing, processing, transmission, indexing, retrieval, and actuation are distributed across networked devices and services. In closely related terminology, the Internet of Audio Things is described as embedding computing and communication capabilities in physical audio-related objects, enabling the reception, processing, and transmission of audio information, while an IoS-centric semantic communication framework describes IoS as combining sound sensing, processing, and transmission techniques to enable collaboration among diverse sound devices (Rovithis et al., 2021, Liang et al., 2024). Across the literature, IoS spans large-scale web audio learning, urban sound sensing and mapping, adaptive soundscape augmentation, provenance and authentication of audio footage, and device-level sonification (Elizalde et al., 2018, Ahmed et al., 2020, Wong et al., 2022, Chenna et al., 2021, Roddy, 23 Sep 2025).
1. Conceptual scope and system layers
A general framework for acoustic sensing systems organizes the field into three layers: a physical layer, a core technique layer, and an application layer. The physical layer includes microphones, speakers, commodity devices, placement and layout, and the characteristics of air-borne and structure-borne acoustic channels. The core technique layer encompasses waveform design, extraction of temporal, spatial, and spectral information, and both model-driven and data-driven methods. The application layer builds on these functions to realize acoustic communication, ranging and localization, radar and imaging, tracking, biometrics, gesture recognition, authentication, touch and force sensing, and environmental sensing (Cai et al., 2019).
Within this layered view, IoS is not limited to passive measurement of sound pressure level. The literature includes continuous web-scale indexing and retrieval of sound, real-time urban sensing, cloud-assisted and fog-assisted classification, adaptive playback systems that augment soundscapes, and embedded sonification that renders non-audio sensor data as sound (Elizalde et al., 2018, Baucas et al., 2019, Wong et al., 2022, Roddy, 23 Sep 2025). This suggests that IoS is best understood as a networked audio systems field rather than as a single sensing application.
A recurrent terminological issue is that adjacent literatures use overlapping acronyms. One strand uses IoS for “Internet of Sounds,” while another uses IoS for “Internet of Senses,” a multisensory telepresence paradigm spanning audio, video, tactile, olfactory, and gustatory streams (Joda et al., 2022). A plausible implication is that precise expansion of the acronym is necessary in cross-disciplinary work.
2. Architectural patterns and compute partitioning
IoS architectures are typically distributed across end devices, intermediate networking or coordination elements, and cloud resources. Representative implementations include cloud-fog-device sensing stacks, serverless edge-to-cloud inference pipelines, and semantic communication systems with sound devices, base stations, SC coordinators, and cloud servers (2002.01376, Wong et al., 2022, Liang et al., 2024).
| System | Edge role | Fog/cloud role |
|---|---|---|
| Real-time noise mapping (Ahmed et al., 2020) | Sound sensor, Arduino Nano, ESP8266 sampling and upload | Firebase storage, Google Maps webapp visualization |
| Urban sound sensing framework (2002.01376) | Recording and local feature extraction on Raspberry Pi 3 Model B | Router as fog node, Ubuntu server for classification |
| Soundscape augmentation (Wong et al., 2022) | Raspberry Pi acquisition and playback, local log-mel computation | AWS IoT Core, Kinesis, Lambda, EFS, MQTT |
| Semantic IoS communication (Liang et al., 2024) | Semantic encoding and local knowledge | SC coordinator, BSs, cloud knowledge and orchestration |
In low-cost sensing deployments, the node often consists of a microphone or sound sensor, a microcontroller, and a wireless module. One real-time urban noise mapping system uses a sound sensor module, ATmega328 (Arduino Nano), and ESP8266 Wi-Fi module, with data uploaded to Firebase and plotted over Google Maps according to sensor location (Ahmed et al., 2020). A separate urban sound sensing framework uses Raspberry Pi 3 Model B units with STM32 NUCLEO-64 and digital MEMS microphones at the device layer, a wireless router as the fog node, and an Ubuntu 16.04 server as the cloud server (2002.01376).
More recent end-to-end systems place substantial intelligence at the edge while retaining cloud aggregation. In residential-area monitoring, wireless acoustic sensor nodes based on Raspberry Pi 3 perform real-time sound classification and direction-of-arrival estimation, then transmit only metadata through WebSocket and MQTT to cloud services and MySQL-backed analytics (Tan et al., 2024). In adaptive soundscape augmentation, the in-situ subsystem computes log-mel spectrograms locally and sends compressed payloads via HTTPS to AWS IoT, while playback units receive cloud-generated augmentation commands via MQTT (Wong et al., 2022).
3. Learning pipelines, sound representations, and semantic indexing
IoS learning pipelines are heterogeneous in both representation and task definition. For urban sound classification, one comparative study evaluates MFCC, Mel-Scale Spectrogram, Chromagram, Spectral Contrast, and Tonnetz features, using a two-hidden-layer neural network with 280 and 300 nodes trained on UrbanSound8K, achieving approximately 85% classification accuracy (Baucas et al., 2019). A related scalable fog framework uses Librosa for on-device feature extraction and a two-layer deep neural network in Tensorflow, again reporting best accuracy at 85% on UrbanSound8K with a learning rate of 0.1 and a 70:30 train-test split (2002.01376).
Residential-area monitoring extends the learning objective from class labels to richer sound metadata. An end-to-end IoT system samples audio at 44.1 kHz, frames data into 8,192 samples per channel, converts raw audio to log-Mel spectrograms, and uses an InceptionNet-based DNN with 335,371 parameters and an approximately 1.3 MB footprint. It classifies 11 urban sound classes and estimates the direction of the dominant sound source using GCC-TDoA, with an angular resolution of 10°. Reported macro/micro-averaged F1 scores are 89%/89% on validation and 84%/89% on test (Tan et al., 2024).
Web-scale IoS introduces a different set of challenges: sound-language mapping, numerous and diverse sound classes, and large-scale evaluation. The Never-Ending Learner of Sounds begins with 600+ sound classes derived from ESC-50, UrbanSound8K, TUT16, and AudioSet; segments crawled web videos into 2.3 seconds; computes 60-band log-mel spectrograms with window size 1024 samples and hop size 512; and applies CNNs for classification. Its indexing and search engine combines sound class predictions, metadata, and user feedback, while Word2Vec and precomputed GloVe embeddings map free-form text queries to known sound classes using cosine similarity with a threshold of 0.15 (Elizalde et al., 2018).
Citizen-science architectures push the same learning logic into participatory settings. A collaborative bird call recognition framework uses mobile devices as audio sensors, deep learning models such as CNN or CRNN on time-frequency representations, GIS-backed soundscape construction, and augmented reality audio interfaces that spatialize virtual bird sounds according to metadata such as location and movement (Rovithis et al., 2021).
4. Monitoring, mapping, and actuation in urban and environmental settings
Urban noise monitoring is a prominent IoS application because it links distributed sensing, geospatial analytics, and policy-relevant outputs. A real-time noise mapping system samples ambient sound every 2 seconds, computes sound intensity in dB, sends JSON packets to Firebase, and plots color-coded overlays on Google Maps. The architecture is designed for multiple node devices installed in different locations of a target area, and the logged data can be used to carry out research regarding sound pollution in targeted areas (Ahmed et al., 2020). The described sound-level formulation associates decibel level with
Mobile nodes extend this model from fixed-point observation to route-based sensing. A scalable, low-cost, real-time environmental noise monitoring system mounts sound nodes on moving vehicles, collects geotagged noise data at 1-second intervals, and uploads them through GSM. The system was deployed in Hyderabad, India, through three measurement campaigns across 27 days, capturing 436,420 data points. Laboratory calibration was shown not to generalize well to mobile data, and calibration based on data collected in a moving environment proved necessary; among the evaluated models, Random Forest Regression achieved the best mobile calibration performance with and RMSE , while incorporating vehicle speed improved performance to , RMSE , and MAE (Manthina et al., 31 Aug 2025).
IoS also includes active manipulation of sonic environments. An adaptive in-situ soundscape augmentation system records ambient audio at 44.1 kHz, computes 64-bin log-mel spectrograms, compresses them into lossless WebP image bitstreams, and uses a serverless cloud pipeline to infer optimal masker tracks and gain settings from 30-second sound snapshots. The cloud side employs AWS Lambda functions and a “probabilistic perceptual attribute predictor” trained on subjective data compliant with ISO 12913-2, drawing candidates from a masker bank of 200 pre-analyzed masker spectrograms. Mean inference runtimes around 8.34 seconds after cold start support near-real-time adaptation, and playback incorporates smooth crossfading (Wong et al., 2022).
Environmental sonification shows a different actuation mode, where the output is not a noise-control masker but a direct auditory rendering of sensor state. One IoS sonification implementation for soil moisture monitoring formalizes on-device wavetable synthesis on a Raspberry Pi Pico, with datalogging, live sonification, and historical sonification modes. Sensor data are mapped both to timbre by interpolation between sine and sawtooth wavetables and to pitch within a G major chord, yielding a dry-soil/high-bright mapping and a wet-soil/lower-dark mapping (Roddy, 23 Sep 2025). This suggests that IoS encompasses both sensing about sound and sound as the interface for other sensed phenomena.
5. Communication semantics, synchronization, and scalability
A central systems question in IoS is where to place feature extraction, classification, storage, and control. Large-scale urban sound classification experiments compare three configurations: edge processing, cloud processing, and a hybrid/fog split in which the end device extracts features and the server performs classification. With Raspberry Pi 3 Model B end devices and up to 12 concurrently active nodes, the hybrid configuration consumes the least power at 1786.86 mW, while the cloud configuration has the fastest average runtime at 16.42 seconds. For latency under scale, the cloud configuration rises sharply to 300.7 ms at 12 devices, whereas edge and hybrid remain low, with the hybrid reaching 5.5 ms at 12 devices; a ranking table therefore places the hybrid configuration highest overall (Baucas et al., 2019).
A closely related fog framework arrives at the same partitioning principle through active low and high power state management and resource reallocation. In the proposed configuration, sound is recorded on the device, feature extraction is also performed on the device, and only the feature vector is sent to the server for classification. This reduces transmission size from approximately 130 KB per raw audio file to 1.6 KB per feature vector, lowers average power consumption to 1786.86 mW, and maintains stable latency up to 12 nodes (2002.01376). The common result is that “all-cloud” and “all-edge” designs are both suboptimal under the studied workloads.
Beyond placement of computation, IoS raises a communication problem specific to sound synchronization. An IoS-centric semantic communication framework identifies three critical synchronization factors: sound quality, timing, and behavior control. Its sender-side pipeline performs sound segmentation, semantic representation, fusion, and symbolic encoding; the transmission stage prioritizes semantic tokens according to their relevance to listeners’ preferences and channel capacity; and the receiver reconstructs and integrates sounds using context- and knowledge-based reasoning (Liang et al., 2024). Under this view, conventional bit-oriented communication is limited because it focuses on bit reproduction rather than the semantic content needed for synchronized listening.
The same framework formalizes semantic importance through knowledge-graph reasoning and probabilistic logic. It also proposes periodic knowledge sharing so that SC models of IoS devices can be updated to optimize synchronization behavior (Liang et al., 2024). A plausible implication is that future IoS communication stacks will be evaluated not only by bitrate and packet error rate but also by semantic adequacy for sound quality, timing, and behavioral adaptation.
6. Provenance, privacy, and unresolved problems
Security in IoS is not confined to channel protection; it also includes provenance of audio content and the authenticity of recordings. A distributed ledger architecture for IoAuT devices combines blockchain with IPFS to store audio recordings and recording-specific metadata, including recTimestamp, deviceMaker, deviceModel, deviceMacAdd, deviceGpsInfo, ipfsHash, contentId, and an acoustic fingerprint recSignature. In a proof-of-concept network of five heterogeneous nodes, the system supports contribution, mining and verification, consumption, and authentication of unknown audio. Fingerprinting experiments show that trimming, amplification, time shift, and pitch shift all alter the signature, enabling detection of edits and forgeries (Chenna et al., 2021).
Privacy concerns appear most directly in field deployments. The residential-area monitoring system transmits only metadata, not raw audio, to the cloud, explicitly supporting privacy requirements while still preserving class index, confidence, SPL, LAeq, timestamp, and DoA for analytics (Tan et al., 2024). At the same time, the blockchain-based authenticity framework notes that metadata are stored in plaintext and that future versions should consider encryption for sensitive fields (Chenna et al., 2021). The resulting tension is structural: provenance benefits from rich metadata, while privacy constraints often favor metadata minimization or protection.
Several open problems recur across the literature. Semantic communication work identifies gaps in mathematical models, adaptive resource allocation, and cross-layer protocol design, especially where local and global semantic knowledge may be inconsistent (Liang et al., 2024). Acoustic sensing research points to waveform and modulation innovation, cross-platform generalization, sensor fusion, privacy and security defenses, and the need for benchmarking and open platforms (Cai et al., 2019). Web-scale learning exposes the continuing difficulty of sound-language mapping, vocabulary growth, and large-scale evaluation without references (Elizalde et al., 2018). Mobile sensing studies show that low-cost sensors require context-matched calibration rather than a one-time laboratory procedure (Manthina et al., 31 Aug 2025). Taken together, these results indicate that IoS remains a systems field in which sensing hardware, machine listening, communication architecture, semantics, and governance are tightly coupled.